[iborakove]

Description

Fixes #13350. Related: #13386 addresses the same issue.

Some models send non-string MCP tool arguments as JSON text, for example {"paths": "[\"a.txt\", \"b.txt\"]"} for an array parameter or "false" for a boolean. coerceArgsToSchema only handled the opposite case (re-stringifying objects for string parameters), so these values reached MCP servers as strings and failed input validation with errors such as expected array, received string.

coerceArgsToSchema (core/tools/parseArgs.ts) now decodes a string argument only when its property schema rejects strings and accepts the decoded JSON type. All other values pass through unchanged, and the MCP server remains responsible for validation.

• Decoding: arrays, objects, booleans, numbers and null are decoded with JSON.parse; "True"/"FALSE" are accepted as getBooleanArg accepts them. Integers beyond Number.MAX_SAFE_INTEGER and non-finite numbers stay strings, so no precision is lost.
• Schema support: type (including type arrays), enum, const, anyOf/oneOf (union), allOf (intersection), boolean schemas and local $refs. Keywords that can only narrow a schema (not, if/then/else, formats, ranges) are ignored, so a type the schema accepts is never excluded.
• String preservation: if the schema accepts strings at all ("string", ["string", "array"], a string branch in anyOf, a string enum/const, or no type constraint), the value is left as is, e.g. "[1]" for a string | array parameter.
• Bounded traversal: analysing one argument's schema visits at most 1,000 schema nodes, expands each $ref once and detects recursive references. Schemas that exceed the budget or recurse leave the argument unchanged, so pathological schemas take bounded time instead of exponential time or a stack overflow.
• Reference safety: only local JSON-pointer references (#, #/…) are followed, through own properties only. References into or across a nested schema resource ($id, or id in draft 4), $dynamicRef/$recursiveRef, and remote or unresolvable references leave the argument unchanged.

Scope: argument normalization in the core tool-call path used by the IDE extensions. The CLI's MCP implementation in extensions/cli is unchanged.

Checklist

• I've read the contributing guide
• The relevant docs, if any, have been updated or created (no user-facing documentation is affected)
• The relevant tests, if any, have been updated or created

Screen recording or screenshot

Not applicable: no UI changes.

Tests

• core/tools/parseArgs.vitest.ts: decoding of each JSON type; valid strings preserved for type, anyOf/oneOf, enum/const and untyped schemas; malformed or mismatched values left unchanged; numeric strings never truncated or rounded; $ref, anyOf, oneOf and allOf resolution; adversarial schemas: an exponential $ref diamond, budget exhaustion (deep nesting, long $ref chains, wide anyOf), recursive references, nested $id resources, $dynamicRef and malformed references.
• core/tools/callTool.vitest.ts (new): the MCP client receives native types (an edits array and a dryRun boolean) across repeated edit_file calls.
• Full core suite: 1,694 tests passed, 0 failed. tsc --noEmit for core and extensions/vscode, ESLint and Prettier pass.
• Manual end-to-end check (not part of this PR) through Continue's MCP client against a strict, schema-validating test MCP server: 30/30 tool calls received correctly typed arguments.